README

README.md

@@ -8,25 +8,97 @@ size_categories:
 - n<1K
 ---
 
-# Workspace-Bench-Lite
+<div align="center">
+  <h1>Workspace-Bench-Lite</h1>
+  <h3>A Lightweight Subset of Workspace-Bench for Fast and Cost-Efficient Evaluation</h3>
+</div>
 
-Workspace-Bench-Lite is the Lite version of Workspace-Bench. It contains 100 tasks and serves as a lightweight subset of the full benchmark, making it suitable for fast evaluation, quick iteration, and low-cost testing.
+<div align="center">
+  <a href="https://workspace-bench.github.io/">
+    <img alt="Project Page" src="https://img.shields.io/badge/Project%20Page-Workspace--Bench-4c6ef5?logo=googlechrome&logoColor=white" />
+  </a>
+  <a href="https://huggingface.co/datasets/Workspace-Bench/Workspace-Bench">
+    <img alt="Full Dataset" src="https://img.shields.io/badge/Hugging%20Face-Full%20Dataset-ff9f1c?logo=huggingface&logoColor=white" />
+  </a>
+  <a href="https://huggingface.co/datasets/Workspace-Bench/Workspace-Bench-Lite">
+    <img alt="Lite Dataset" src="https://img.shields.io/badge/Hugging%20Face-Lite%20Dataset-f76707?logo=huggingface&logoColor=white" />
+  </a>
+  <a href="https://arxiv.org/abs/2605.03596">
+    <img alt="arXiv Paper" src="https://img.shields.io/badge/arXiv-2605.03596-b31b1b?logo=arxiv&logoColor=white" />
+  </a>
+</div>
+
+<div align="center">
+  <a href="#overview">Overview</a> •
+  <a href="#leaderboard">LeaderBoard</a> •
+  <a href="#distribution">Distribution</a> •
+  <a href="#quick-start">Quick Start</a> •
+  <a href="#citation">Citation</a>
+</div>
 
 ## Overview
 
-- 100 tasks
-- Lightweight version of Workspace-Bench
-- Suitable for quick testing and rapid benchmarking
-- Preserves the benchmark's core workspace-learning setting
+Workspace-Bench-Lite is the **Lite split of Workspace-Bench 1.0**, designed for fast iteration and lower-cost benchmarking while preserving the core evaluation setting of the full benchmark.
+
+It contains **100 tasks** selected from the full Workspace-Bench and is intended to maintain distributional fidelity across:
+
+- workspace roles
+- task difficulty levels
+- workspace-learning capabilities
+- file dependency patterns
+
+Workspace-Bench-Lite keeps the central benchmark objective unchanged: evaluating whether AI agents can perform **Workspace Learning**, namely identifying, reasoning over, and acting on explicit and implicit dependencies among heterogeneous files in realistic workspaces.
+
+Compared with the full benchmark, Workspace-Bench-Lite is more suitable for:
+
+- rapid prototyping of agent systems
+- low-cost model and harness comparison
+- regression testing during agent development
+- lightweight public leaderboard evaluation
+
+<div align="center">
+  <img src="assets/Framework.png" alt="Workspace-Bench-Lite framework overview" width="980" />
+</div>
+
+The figure above summarizes the overall Workspace-Bench framework that Lite inherits from the full benchmark: role-specific workspaces, cross-file dependent tasks, and capability-oriented rubric evaluation. Workspace-Bench-Lite keeps this same evaluation setting while offering a more lightweight split for faster experimentation.
+
+In the Workspace-Bench paper, the Lite split is described as a **100-task subset** that preserves the benchmark distribution while reducing evaluation costs by about **70%**. This makes it a practical entry point for researchers and developers who want to evaluate workspace agents without running the full 388-task benchmark.
+
+The Lite split remains grounded in the same benchmark philosophy as the full version:
 
-## Links
+- realistic workspace tasks rather than isolated QA
+- heterogeneous files rather than single-format inputs
+- dependency-aware task solving rather than direct prompt lookup
+- rubric-based evaluation rather than only final-answer matching
+
+## LeaderBoard
+
+<div align="center">
+  <img src="assets/rubrics_success.png" alt="Workspace-Bench-Lite leaderboard snapshot" width="980" />
+</div>
+
+The figure above shows a leaderboard snapshot of rubric pass rates on Workspace-Bench-Lite across different agent harnesses and backbone models. It highlights the same central takeaway as the full benchmark: current agents still show a clear gap from human-level performance, and both the foundation model and the harness design materially affect final results.
+
+## Distribution
+
+<div align="center">
+  <img src="assets/Distribution.png" alt="Workspace-Bench-Lite distribution overview" width="980" />
+</div>
+
+The distribution figure provides a compact overview of the Lite split, including file types, task abilities, task difficulty, workspace allocation, rubric counts, required files per task, and dependency structure. Although lighter than the full benchmark, Workspace-Bench-Lite is designed to preserve the key distributional characteristics of Workspace-Bench.
+
+## Quick Start
+
+Please refer to the official GitHub repository for benchmark setup, evaluation instructions, and updates:
 
-- Official website: https://workspace-bench.github.io/
 - GitHub repository: https://github.com/OpenDataBox/Workspace-Bench
-- arXiv paper: https://arxiv.org/abs/2605.03596
+
+More Lite-specific usage notes will be added in future updates.
 
 ## Citation
 
+If you use Workspace-Bench-Lite in your research, please cite the Workspace-Bench paper:
+
 ```bibtex
 @misc{tang2026workspacebench10benchmarkingai,
   title={Workspace-Bench 1.0: Benchmarking AI Agents on Workspace Tasks with Large-Scale File Dependencies},